Multi-path Scheduling with Deep Reinforcement Learning

Abstract

We present a practical approach of how deep learning models can improve 5G network service. We demonstrate the potential of a deep Q-network agent applied to a traffic management problem, consisting in the path selection in a multi-path scenario. We use for the demonstration a multi-path QUIC implementation and we train an agent for improving the algorithm that selects the optimal path, with results in a better utilization of the network by increasing the aggregated throughput of the multi-path flows, as we detail in the results of this work.